Pith. sign in

REVIEW 4 major objections 4 minor 40 references

CRAB: Camera-Radar Fusion for Reducing Depth Ambiguity in Backward Projection based View Transformation

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read CRAB claims that fusing frustum-view radar occupancy with image depth distributions resolves the same-ray depth ambiguity of backward projection, reaching 62.4% NDS and 54.0% mAP on nuScenes.

desk verdict Solid camera-radar fusion engineering with strong nuScenes numbers, but the ablation doesn't isolate the occupancy prior and the paper has reproducibility warts; worth reviewing nonetheless. read the letter →

arxiv 2509.05785 v1 pith:HBAAAP64 submitted 2025-09-06 cs.CV

classification cs.CV
keywords camera-radarfusion3Dobjectdetectiondepthambiguitybackwardprojectionbird'seyeviewradaroccupancydeformableattentionnuScenes
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CRAB is a camera-radar 3D detector targeting a specific failure in backward-projection view transformation: BEV queries lying on the same camera ray retrieve the same image feature, so objects at different depths blur together and produce duplicate false positives. The paper's central claim is that radar's sparse but precise depth can resolve that ambiguity when fused multiplicatively with the image's dense but unreliable depth information. It reports 62.4% NDS and 54.0% mAP on nuScenes, the best among backward-projection camera-radar fusion methods, and shows the same BEV features also improve segmentation over its camera baseline. A reader should care because the result suggests that, on this benchmark, a low-cost camera-radar setup can approach the geometric quality of LiDAR-based detectors without the cost and weather sensitivity of LiDAR.

What carries the argument

Two objects carry the argument. The first is the joint occupancy tensor O_IR = I_O ⊗ R_O, the outer product of perspective-view image occupancy (dense, per-pixel, depth-unreliable) and frustum-view radar occupancy (sparse, per-depth-bin, precise). Its role is to gate where image context is placed so same-ray queries separate by depth. The second is the frustum-shaped radar context map R_C(u,d) with height 1, the target of RCSCA's deformable attention; it lets the BEV query fetch RCS and Doppler features by exact depth position while staying in the same frustum coordinate frame used by the image stream.

What would settle it

Replace R_O in Eq. (2) with an all-ones tensor of the same shape and re-run the nuScenes ablation: if NDS and mAP stay near 58.5/49.9, radar occupancy is not doing the claimed depth disambiguation. As a second check, halve radar sweep accumulation or randomly drop 50% of points and look for the reappearance of same-ray duplicate detections and a fall toward the DFA3D baseline.

Watch

Extended reading notes

Core claim

The paper claims that depth ambiguity is the main defect of backward projection, and that it can be removed by giving each BEV query a radar-validated depth. Its first module, ROSCA, forms a 3D occupancy tensor as the outer product of image occupancy and frustum-view radar occupancy, then multiplies in image context; because radar occupancy is precise per depth bin, queries on the same ray sample different depth positions and receive different features. Its second module, RCSCA, projects the image-encoded BEV query onto a height-1 frustum-shaped radar context map and applies deformable attention, so each query also collects radar cross-section and Doppler information from its own (u,d) locat

Load-bearing premise

The chain rests on radar occupancy being accurate and dense enough after voxelization that multiplying it with image occupancy steers image features to the correct depth bin; if radar points are too sparse or too noisy, that multiplication would suppress or misplace most image context and the reported gains would not transfer.

Editorial extensions

If this is right

  • Same-ray BEV queries no longer collapse to identical image features, which the qualitative results show as fewer duplicate detections around ego-centered rays.
  • Radar context (RCS and Doppler velocity) is load-bearing for the final score: adding RCSCA on top of ROSCA raises NDS from 57.0 to 58.5 and mAP from 48.7 to 49.9 in the single-scale ablation.
  • Keeping radar context in frustum view rather than BEV is worth 1.3 NDS and 1.5 mAP, because the attention operates in the same coordinate frame as the image stream.
  • The fused BEV representation transfers to a second task: segmentation mIoU over the camera baseline BEVFormer improves from 45.6 to 48.5, with gains in every class.
  • Radar's weather robustness carries through: CRAB beats the camera-only baseline and other fusion models in rain and at night, and at night it also beats LiDAR-only CenterPoint-P (40.0 vs 34.7 NDS).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same outer-product gating could be inserted into other lifted-feature backward-projection detectors, so the depth-ambiguity fix may transfer beyond the DFA3D-style baseline; the authors mention applying it to other backward-projection methods only as future work.
  • Because RCSCA carries Doppler velocity into the BEV feature, the representation is naturally suited to motion prediction and trajectory forecasting, tasks the paper does not evaluate but where radial velocity is a direct cue.
  • A stress test the paper does not run is radar degradation: artificially thinning the point cloud or injecting multipath noise would show how much of the gain is occupancy accuracy versus sheer sensor redundancy; ablating R_O to all-ones would isolate the multiplicative fusion's contribution.
  • The reported gap to LiDAR-only CenterPoint-V on validation (62.4 vs 65.3 NDS) suggests the practical ceiling is close, but that comparison is benchmark-specific; whether it holds in other weather, sensor setups, or dense urban scenes is untested.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. CRAB proposes a camera-radar fusion architecture for 3D object detection and BEV segmentation built on backward-projection view transformation. Its first module, ROSCA, multiplies image occupancy with frustum-view radar occupancy to produce a 3D occupancy that weights image context along depth, and its second module, RCSCA, applies deformable cross-attention between BEV queries and a frustum-shaped radar context feature map. Evaluated on nuScenes, the method reports 62.4% NDS and 54.0% mAP on the test set, which the authors claim is state-of-the-art among backward-projection camera-radar fusion methods. The paper includes component ablations, a frustum-vs-BEV analysis, weather/lighting robustness results, and qualitative visualizations.

Significance. If the causal claims are supported, the paper makes a useful contribution: it adapts the backward-projection paradigm (DFA3D/BEVFormer) to incorporate radar as a depth prior, which is a natural and previously underexplored combination. The two-module design is simple, the evaluation covers both detection and segmentation on a public benchmark, and the weather/lighting analysis (Table VI) is a positive addition. The paper gives explicit credit to prior work, uses public baselines, and includes ablations. However, the central attribution of the gains to the radar-occupancy depth prior is not isolated by the provided ablations, the ablations are run only on a reduced configuration, and the main comparison table contains a clear data error. The contribution is therefore promising but needs additional evidence before the stated mechanism can be accepted.

major comments (4)
  1. [Sec. IV-D.1 / Table IV] The ablation does not isolate the claimed depth-prior mechanism. From (A)-(D): ROSCA alone adds 6.9 NDS (50.1→57.0), RCSCA alone adds 7.9 NDS (50.1→58.0), and the combination adds 8.4 NDS. The marginal gain of ROSCA over RCSCA is therefore only 0.5 NDS, whereas RCSCA alone accounts for most of the total improvement. Since RCSCA also gives same-ray queries different depth coordinates along (u,d) (Sec. III-C.2), the numbers are at least as consistent with the alternative that the gain comes from radar context (RCS/Doppler) rather than from the Eq. (2) occupancy prior that the abstract names as the core novelty. Please provide an ablation that controls for this, e.g., ROSCA with a non-depth or randomized radar occupancy, or a matched comparison with RCSCA removed.
  2. [Sec. IV-B / Table IV vs Table I] The component analysis is run with 3 encoder layers and a 150×150 BEV grid, while the headline results in Tables I and II use 6 layers and 200×200. The Table IV baseline (BEVFormer-S-DFA3D, 50.1 NDS) is not the full DFA3D baseline reported in Table I (53.1 NDS). Consequently, the modular contributions are established only on a reduced-capacity configuration, and it is not shown that they transfer to the setting used for the SOTA claim. Report the same ablation on the final configuration, or justify the transfer with additional evidence.
  3. [Sec. III-C.1 / Eq. (2) / Abstract] There is a mismatch between the stated mechanism and the equations. The text says the method combines 'the dense but unreliable depth distribution from images' with radar occupancy, but Eq. (2) multiplies the image occupancy I_O with radar occupancy R_O; the depth distribution I_D is used only in the separate branch B_I and is fused later in Eq. (3). As written, the abstract and Sec. I attribute the depth-distinction improvement to a mechanism that the architecture does not implement in the named equation. Please revise the description to match the actual computation and justify why image occupancy, not depth distribution, is the right quantity to multiply with radar occupancy.
  4. [Table I] The row for RCBEV4d (Modality C&R, Backbone Swin, V.T. Forward) is numerically identical to the TransCAR row above it (NDS 49.7, mAP 38.1, and all five error metrics). This appears to be a copy-paste error. The correct RCBEV4d results must be supplied and the table regenerated.
minor comments (4)
  1. [Eq. (2)] The symbol ⊗ is described as an outer product, but the dimensions in Eq. (2) imply broadcast multiplication with axis expansion. Please define the operation explicitly to avoid confusion with the standard tensor outer product.
  2. [Throughout] Typographical issues: 'concatentation' in Sec. III-C.1, 'V ovNet' in Table I, '2×10 −4' in Sec. IV-B, and extra spaces before commas in phrases like 'BEV ,'.
  3. [Sec. IV-D.3 / Table VI] The weather and lighting splits are not described. Report the number of scenes per condition and any class imbalance, since the comparisons in Table VI may be sensitive to scene composition.
  4. [Table II] The test-set comparison uses different backbones across baselines (e.g., RCM-Fusion with ResNet101, TransCAR with V2-99, CRAB with V2-99). A same-backbone comparison on the test set, or at least on the validation set, would strengthen the state-of-the-art claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CRAB's reported improvements are measured against external nuScenes benchmarks, and its modules are new combinations of existing components; self-citations are implementation details, not load-bearing evidence.

full rationale

The central claim is empirical: CRAB's 62.4% NDS and 54.0% mAP are obtained by training on the public nuScenes dataset and comparing against external baselines. No equation reduces a claimed prediction to a fitted parameter; Eq. 2 defines the ROSCA computation, but the reported metrics come from held-out evaluation, not from the equation itself. The paper's self-citations to CRN [8] and GRIF Net [5] are used for implementation details (frustum-voxel radar feature extraction and radar sweep accumulation) rather than as proof of the SOTA result, so they are not load-bearing circularity. The ablation in Table IV does raise a causal-attribution question—ROSCA's marginal contribution over RCSCA is small (+0.5 NDS)—but that is a concern about whether the named depth-ambiguity mechanism is the true source of gain, not about circularity. The paper itself labels part of the explanation as a conjecture, which further confirms it is not presenting a formal derivation. No uniqueness theorem, ansatz, or definitional equivalence is imported from prior work to force the outcome. The design is self-contained against external benchmarks and the reported improvements are not equivalent to the inputs by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim is empirical and rests on learned network weights plus manual hyperparameters; there are no fitted physical constants. The load-bearing assumptions are sensor and calibration accuracy, the validity of the nuScenes benchmark, and the multiplicative fusion assumption in ROSCA.

free parameters (4)
  • BEV grid size = 200x200 for Tables I, II, VI; 150x150 for Tables III, IV, V
    Chosen by hand; directly affects spatial resolution and performance, and the paper switches between values across experiments.
  • Number of encoder layers = 6 for Tables I, II, VI; 3 for Tables III, IV, V
    Manual architectural choice; the difference complicates direct comparison of the main results and the ablation results.
  • Radar sweep accumulation = 6 previous sweeps
    Follows GRIF Net [5]; affects radar occupancy density and is not ablated in the paper.
  • Loss weights = 1, 1, 1 in Eq. (5)
    The total loss is a plain sum of detection, depth, and occupancy losses; the relative weighting is a free design choice not analyzed.
assumptions (5)
  • domain assumption Radar provides accurate depth and is robust in adverse weather, so it can serve as a sparse but precise depth prior.
    Motivates the design of ROSCA in Sec. III-C.1; if radar depth noise is high, the occupancy prior is wrong.
  • domain assumption Camera-radar extrinsic and intrinsic calibration is accurate.
    Radar points are projected onto perspective images and voxelized into frustum view in Sec. III-B; calibration errors would misalign the occupancy.
  • domain assumption nuScenes ground truth and metrics reliably measure 3D perception quality.
    All conclusions are drawn from nuScenes validation and test sets, Sec. IV-A.
  • standard math BEVFormer, DFA3D, and deformable attention are valid building blocks.
    Eqs. (1) through (4) rely on the attention and projection mechanisms introduced in [14], [15], and [16].
  • ad hoc to paper Image occupancy and radar occupancy can be combined multiplicatively into a joint 3D occupancy.
    Eq. (2) defines O_IR = I_O ⊗ R_O; this multiplicative fusion is proposed without derivation or independent evidence beyond the reported ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CRAB: Camera-Radar Fusion for Reducing Depth Ambiguity in Backward Projection based View Transformation." pith.science (2026). https://pith.science/paper/HBAAAP64

@misc{pith2026250905785,
  author       = {Pith},
  title        = {Pith review of: CRAB: Camera-Radar Fusion for Reducing Depth Ambiguity in Backward Projection based View Transformation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HBAAAP64}},
  note         = {Machine review of arXiv:2509.05785}
}
read the original abstract

Recently, camera-radar fusion-based 3D object detection methods in bird's eye view (BEV) have gained attention due to the complementary characteristics and cost-effectiveness of these sensors. Previous approaches using forward projection struggle with sparse BEV feature generation, while those employing backward projection overlook depth ambiguity, leading to false positives. In this paper, to address the aforementioned limitations, we propose a novel camera-radar fusion-based 3D object detection and segmentation model named CRAB (Camera-Radar fusion for reducing depth Ambiguity in Backward projection-based view transformation), using a backward projection that leverages radar to mitigate depth ambiguity. During the view transformation, CRAB aggregates perspective view image context features into BEV queries. It improves depth distinction among queries along the same ray by combining the dense but unreliable depth distribution from images with the sparse yet precise depth information from radar occupancy. We further introduce spatial cross-attention with a feature map containing radar context information to enhance the comprehension of the 3D scene. When evaluated on the nuScenes open dataset, our proposed approach achieves a state-of-the-art performance among backward projection-based camera-radar fusion methods with 62.4\% NDS and 54.0\% mAP in 3D object detection.

Figures

Figures reproduced from arXiv: 2509.05785 by the authors.

Figure 1
Figure 1. Comparison of camera-radar fusion methods. (a) In CRAFT, 3D proposals are generated directly from the perspective view image using a detector, and fusion involves filtering noisy radar points outside the proposals and 2D deformable attention. (b) CRN utilizes forward projection, incorporating radar occupancy in the ‘lift’ stage, and adaptively fuses image and radar BEV features with 2D deformable attention without r… view at source ↗
Figure 2
Figure 2. Overall architecture of CRAB. CRAB extracts image and radar features from their respective backbones. Afterwards, BEV features obtained through Depth Distribution guided Spatial Cross Attention and (a) Radar Occupancy guided Spatial Cross Attention are fused together. The fused BEV feature and radar context features undergo (b)spatial cross attention, before finally passing through a task-specific heads. deformable … view at source ↗
Figure 4
Figure 4. Comparison of Encoded BEV feature. During view transforma￾tion, (a) does not consider depth, resulting features in red triangular area(ray area) being somewhat scattered and indistinct. In contrast, the BEV feature of the proposed architecture (b) shows a clear distinction in the depth of features, demonstrating improved clarity. After passing through the two modules presented earlier, as shown in [PITH_FULL_IMAGE:… view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: Illustration of proposed Radar Occupancy-guided Spatial Cross Attention. The 3D occupancy generated by the outer product of radar occupancy and image occupancy, along with the outer product of image context feature, is used to obtain an expanded feature. Through spatia…
Figure 6
Figure 6. Figure 6: Qualitative results of BEV segmentation on nuScenes val set. Our proposed method demonstrates the ability to construct maps accurately in complex scenarios. Best viewed in color with zoom in. TABLE VI ANALYSIS OF DIFFERENT LIGHTING AND WEATHER CONDITIONS. BOLD AND UNDE…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 37 canonical work pages

  1. [1]

    Supfusion: Supervised lidar-camera fusion for 3d object detection,

    Y . Qin, C. Wang, Z. Kang, N. Ma, Z. Li, and R. Zhang, “Supfusion: Supervised lidar-camera fusion for 3d object detection,” inProc. IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp. 22 014–22 024

  2. [2]

    Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,

    Z. Liu, H. Tang, A. Amini, X. Yang, H. Mao, D. Rus, and S. Han, “Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,” inProc. IEEE International Conference on Robotics and Automation (ICRA), 2023

  3. [3]

    Cross modal transformer: Towards fast and robust 3d object detection,

    J. Yan, Y . Liu, J. Sun, F. Jia, S. Li, T. Wang, and X. Zhang, “Cross modal transformer: Towards fast and robust 3d object detection,” inProc. IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp. 18 268–18 278

  4. [4]

    Autoalignv2: Deformable feature aggregation for dynamic multi-modal 3d object detection,

    Z. Chen, Z. Li, S. Zhang, L. Fang, Q. Jiang, and F. Zhao, “Autoalignv2: Deformable feature aggregation for dynamic multi-modal 3d object detection,” inProc. European Conference on Computer Vision, 2022

  5. [5]

    Grif net: Gated region of interest fusion network for robust 3d object detection from radar point cloud and monocular image,

    Y . Kim, J. W. Choi, and D. Kum, “Grif net: Gated region of interest fusion network for robust 3d object detection from radar point cloud and monocular image,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, pp. 10 857–10 864

  6. [6]

    Craft: Camera-radar 3d object detection with spatio-contextual fusion transformer,

    Y . Kim, S. Kim, J. W. Choi, and D. Kum, “Craft: Camera-radar 3d object detection with spatio-contextual fusion transformer,” inProc. AAAI Conference on Artificial Intelligence (AAAI), 2022, pp. 1160– 1168

  7. [7]

    Centerfusion: Center-based radar and camera fusion for 3d object detection,

    R. Nabati and H. Qi, “Centerfusion: Center-based radar and camera fusion for 3d object detection,” inProc. IEEE Winter Conference on Applications of Computer Vision (WACV), 2021

  8. [8]

    Crn: Camera radar net for accurate, robust, efficient 3d perception,

    Y . Kim, J. Shin, S. Kim, I.-J. Lee, J. W. Choi, and D. Kum, “Crn: Camera radar net for accurate, robust, efficient 3d perception,” inProc. IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 17 615–17 626

Show all 40 references
  1. [9]

    Rcm-fusion: Radar-camera multi-level fusion for 3d object detection,

    J. Kim, M. Seong, G. Bang, D. Kum, and J. W. Choi, “Rcm-fusion: Radar-camera multi-level fusion for 3d object detection,” 2023

  2. [10]

    Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection,

    T. Zhou, J. Chen, Y . Shi, K. Jiang, M. Yang, and D. Yang, “Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection,”IEEE Transactions on Intelligent Vehicles, vol. 8, no. 2, pp. 1523–1535, 2023

  3. [11]

    Transcar: Transformer-based camera-and-radar fusion for 3d object detection,

    S. Pang, D. Morris, and H. Radha, “Transcar: Transformer-based camera-and-radar fusion for 3d object detection,” inProc. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023, pp. 10 902–10 909

  4. [12]

    Redformer: Radar enlightens the darkness of camera perception with transformers,

    C. Cui, Y . Ma, J. Lu, and Z. Wang, “Redformer: Radar enlightens the darkness of camera perception with transformers,”IEEE Transactions on Intelligent Vehicles, vol. 9, no. 1, pp. 1358–1368, 2024

  5. [13]

    Fb-bev: Bev representation from forward-backward view transfor- mations,

    Z. Li, Z. Yu, W. Wang, A. Anandkumar, T. Lu, and J. M. Alvarez, “Fb-bev: Bev representation from forward-backward view transfor- mations,” inProc. IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 6919–6928

  6. [14]

    Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers,

    Z. Li, W. Wang, H. Li, E. Xie, C. Sima, T. Lu, Y . Qiao, and J. Dai, “Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers,” inProc. Eu- ropean Conference on Computer Vision, 2022, pp. 1–18

  7. [15]

    Dfa3d: 3d deformable attention for 2d-to-3d feature lifting,

    H. Li, H. Zhang, Z. Zeng, S. Liu, F. Li, T. Ren, and L. Zhang, “Dfa3d: 3d deformable attention for 2d-to-3d feature lifting,” inProc. IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 6684–6693

  8. [16]

    Deformable detr: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to-end object detection,” 2021

  9. [17]

    nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” inProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

  10. [18]

    Mvfusion: Multi-view 3d object detection with semantic-aligned radar and camera fusion,

    Z. Wu, G. Chen, Y . Gan, L. Wang, and J. Pu, “Mvfusion: Multi-view 3d object detection with semantic-aligned radar and camera fusion,” inProc. IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 2766–2773

  11. [19]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

  12. [20]

    An energy and gpu-computation efficient backbone network for real-time object detection,

    Y . Lee, J.-w. Hwang, S. Lee, Y . Bae, and J. Park, “An energy and gpu-computation efficient backbone network for real-time object detection,” inProc. IEEE Conference on Computer Vision and Pattern Recognition Workshops(CVPRW), 2019

  13. [21]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” inProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 936–944

  14. [22]

    Deep ordi- nal regression network for monocular depth estimation,

    H. Fu, M. Gong, C. Wang, K. Batmanghelich, and D. Tao, “Deep ordi- nal regression network for monocular depth estimation,” inProc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

  15. [23]

    Bevdepth: Acquisition of reliable depth for multi-view 3d object detection,

    Y . Li, Z. Ge, G. Yu, J. Yang, Z. Wang, Y . Shi, J. Sun, and Z. Li, “Bevdepth: Acquisition of reliable depth for multi-view 3d object detection,” inProc. AAAI Conference on Artificial Intelligence (AAAI), vol. 37, 2023, pp. 1477–1485

  16. [24]

    Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,

    J. Philion and S. Fidler, “Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,” inProc. European Conference on Computer Vision, 2020

  17. [25]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” inProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

  18. [26]

    Pointnet++: Deep hierar- chical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierar- chical feature learning on point sets in a metric space,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

  19. [27]

    Second: Sparsely embedded convolutional detection,

    Y . Yan, Y . Mao, and B. Li, “Second: Sparsely embedded convolutional detection,”Sensors, vol. 18, no. 10, pp. 3337–3352, 2018

  20. [28]

    MMDetection3D: OpenMMLab next-generation platform for general 3D object detection,

    M. Contributors, “MMDetection3D: OpenMMLab next-generation platform for general 3D object detection,” https://github.com/open- mmlab/mmdetection3d, 2020

  21. [29]

    Center-based 3d object detection and tracking,

    T. Yin, X. Zhou, and P. Krahenbuhl, “Center-based 3d object detection and tracking,” inProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021, pp. 11 784–11 793

  22. [30]

    Matrixvt: Efficient multi- camera to bev transformation for 3d perception,

    H. Zhou, Z. Ge, Z. Li, and X. Zhang, “Matrixvt: Efficient multi- camera to bev transformation for 3d perception,” inProc. IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 8548– 8557

  23. [31]

    Class-balanced grouping and sampling for point cloud 3d object detection,

    B. Zhu, Z. Jiang, X. Zhou, Z. Li, and G. Yu, “Class-balanced grouping and sampling for point cloud 3d object detection,” 2019. [Online]. Available: https://arxiv.org/abs/1908.09492

  24. [32]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. [Online]. Available: https://openreview.net/forum?id=Bkg6RiCqY7

  25. [33]

    Hdmapnet: An online hd map construction and evaluation framework,

    Q. Li, Y . Wang, Y . Wang, and H. Zhao, “Hdmapnet: An online hd map construction and evaluation framework,” 2022

  26. [34]

    End-to-end vectorized hd- map construction with piecewise bezier curve,

    L. Qiao, W. Ding, X. Qiu, and C. Zhang, “End-to-end vectorized hd- map construction with piecewise bezier curve,” inProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 13 218–13 228

  27. [35]

    Bevsegformer: Bird’s eye view semantic segmentation from arbitrary camera rigs,

    L. Peng, Z. Chen, Z. Fu, P. Liang, and E. Cheng, “Bevsegformer: Bird’s eye view semantic segmentation from arbitrary camera rigs,” inProc. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), January 2023, pp. 5935–5943

  28. [36]

    Online map vectorization for autonomous driving: A rasterization perspective,

    G. Zhang, J. Lin, S. Wu, y. song, Z. Luo, Y . Xue, S. Lu, and Z. Wang, “Online map vectorization for autonomous driving: A rasterization perspective,” inAdvances in Neural Information Processing Systems, 2023, pp. 31 865–31 877

  29. [37]

    Detr3d: 3d object detection from multi-view images via 3d- to-2d queries,

    Y . Wang, V . Guizilini, T. Zhang, Y . Wang, H. Zhao, , and J. M. Solomon, “Detr3d: 3d object detection from multi-view images via 3d- to-2d queries,” inThe Conference on Robot Learning (CoRL), 2021

  30. [38]

    Bevformer v2: Adapting modern image backbones to bird’s-eye-view recognition via perspec- tive supervision,

    C. Yang, Y . Chen, H. Tian, C. Tao, X. Zhu, Z. Zhang, G. Huang, H. Li, Y . Qiao, L. Lu, J. Zhou, and J. Dai, “Bevformer v2: Adapting modern image backbones to bird’s-eye-view recognition via perspec- tive supervision,” inProc. IEEE/CVF Conference on Computer Vision and Pattern...

  31. [39]

    Polarformer: Multi-camera 3d object detection with polar transformer,

    Y . Jiang, L. Zhang, Z. Miao, X. Zhu, J. Gao, W. Hu, and Y .-G. Jiang, “Polarformer: Multi-camera 3d object detection with polar transformer,” inProc. AAAI Conference on Artificial Intelligence, 2023, pp. 1042–1050

  32. [40]

    Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving,

    Y . Wei, L. Zhao, W. Zheng, Z. Zhu, J. Zhou, and J. Lu, “Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp. 21 729–21 740

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.